Dossier

What dossier MCP server does

The dossier MCP server exposes the Dossier workflow to an MCP-compatible AI client. It works with .ds.md files: Markdown-based agent skills containing JSON frontmatter such as a title, version, protocol version, status, objective, and risk level. The format adds version pinning, checksums, and Ed25519 signatures to skills that an AI agent can read and execute.

Dossiers may be stored in a GitHub repository, on an organization’s own server, or in the Dossier registry. Their location does not change the verification format. A client can discover available dossiers, check their integrity and authenticity, and run a selected workflow without manually copying a URL into a conversation.

The dossier MCP server is intended for model-driven orchestration. The model chooses steps, tool calls, retries, and completion conditions from the skill instructions, while Dossier provides a deterministic scheduler around multi-step runs. The scheduler itself does not call an LLM.

How it works

A dossier starts as a .ds.md file written in Markdown with a ---dossier JSON frontmatter block. The author signs the skill and associates it with a specific version. Before execution, Dossier can check the file’s checksum and signature so users can confirm that it matches the signed content and identify its author.

After verification, the AI agent follows the skill’s instructions and evaluates its validation criteria. This creates a hybrid approach: orchestration inside a skill is model-driven, while the surrounding scheduling behavior is deterministic. The signature establishes integrity and origin, but it does not protect against prompt injection or prove that the instructions are safe to run.

Dossier is also described as a layer over Agent Skills and SKILL.md, rather than a replacement for that format. Skills can be moved into or out of the Dossier format with the project’s CLI commands.

Setup and configuration

For Claude Code, add the server with:

claude mcp add dossier --scope user -- npx @ai-dossier/mcp-server

A manual MCP configuration can run the package with npx:

{
  "mcpServers": {
    "dossier": {
      "command": "npx",
      "args": ["-y", "@ai-dossier/mcp-server"]
    }
  }
}

The manual configuration is shown for Claude Desktop and other MCP clients. The repository also documents a Claude Code plugin installation. Creating or authoring dossiers uses the separate CLI package: npx @ai-dossier/cli init initializes the local Dossier setup, and npx @ai-dossier/cli create my-workflow scaffolds a dossier.

Tools and capabilities

The dossier MCP server supports these documented workflows:

• Discover available dossiers.
• Verify dossier checksums and signatures.
• Execute selected dossier workflows.
• Work with dossiers hosted in GitHub, on another server, or in the Dossier registry.
• Use signed and versioned skills across compatible AI tools.

Dossier files use .ds.md for immutable instructions and .dsw.md for mutable working files. The protocol specification is listed as version 1.0 in the repository, and the CLI and MCP server are distributed as npm packages.

Limitations and notes

Verification confirms integrity and origin; it does not make a dossier safe, stop prompt injection, or replace review of the instructions. Execution remains model-driven within each skill, so the model determines the operational steps. The deterministic scheduler surrounds the run but never calls an LLM. The provided setup examples use npx, so the client environment must be able to invoke that command and reach the selected dossier location.

The dossier MCP server is therefore a fit for teams that want portable agent workflows with explicit versions and verification, not for workflows that require signatures alone to provide security guarantees.